First of all, it seems that you are using the wrong hash mode, i.e. the algorithm is (or at least seems to be) salted MD5 e.g. -m 10 = md5 ($pass . $salt) and not raw-md5 (-m 0 = MD5).
The speed difference should be very clear after reading e.g. this https://hashcat.net/wiki/frequently_asked_questions but also hundreds of other posts that asked exactly the same question. The most important thing you need to know about is "acceleration" and how it works (and when exactly it is possible or when exactly we can accelerate the most).
If you just run a dictionary without rules, the GPU can't accelerate the cracking a lot (if at all), in some cases with a recent and fast CPU straight attacks can be even faster with cpu hashcat.
What you discovered also is that the left-hand side of the mask shouldn't be too small, this is of course true. The reason is that oclHashcat changes the left-hand side most frequently (and if the left-hand side is small and even has some hard-coded strings, acceleration will be much less than "normal" i.e. with a mask of e.g. ?a?a?a?a at the very beginning).
Several of these good-to-know facts about hashcat/oclHashcat are explained in the documentation (wiki / faq / forum etc), I'm sure you will find out some interesting answers and ideas about how to best crack hashes. For instance, the same way that you can use -m 10 = md5 ($pass . $salt) to append a static string, you can also use -m 20 = md5 ($salt . $pass) to prepend a static string (e.g. the "0.9" in your case can be seen as a static "left prepended salt". I'm not saying that it is always faster to switch from -m 10 (or -m 0) to -m 20 if possible, but it some cases it could at least be worth a try. There are many more such tricks you can discover by just reading the (*****) manual.
The speed difference should be very clear after reading e.g. this https://hashcat.net/wiki/frequently_asked_questions but also hundreds of other posts that asked exactly the same question. The most important thing you need to know about is "acceleration" and how it works (and when exactly it is possible or when exactly we can accelerate the most).
If you just run a dictionary without rules, the GPU can't accelerate the cracking a lot (if at all), in some cases with a recent and fast CPU straight attacks can be even faster with cpu hashcat.
What you discovered also is that the left-hand side of the mask shouldn't be too small, this is of course true. The reason is that oclHashcat changes the left-hand side most frequently (and if the left-hand side is small and even has some hard-coded strings, acceleration will be much less than "normal" i.e. with a mask of e.g. ?a?a?a?a at the very beginning).
Several of these good-to-know facts about hashcat/oclHashcat are explained in the documentation (wiki / faq / forum etc), I'm sure you will find out some interesting answers and ideas about how to best crack hashes. For instance, the same way that you can use -m 10 = md5 ($pass . $salt) to append a static string, you can also use -m 20 = md5 ($salt . $pass) to prepend a static string (e.g. the "0.9" in your case can be seen as a static "left prepended salt". I'm not saying that it is always faster to switch from -m 10 (or -m 0) to -m 20 if possible, but it some cases it could at least be worth a try. There are many more such tricks you can discover by just reading the (*****) manual.